> "I think this is the wrong way round, there should be an URL/URI there and a formatter to generate a local string for display purposes." -JB

Where this URL/URI-to-display is possible, I agree with Jerven; however I'm (so far) completely naive about wikidata's ingest process (apologies). 
Sometimes that accession is all you've got to work with at the start. It is in some third-party'es database cross reference tables and that's pretty much it.

Even worse, sometimes the type of the encountered accession is implied by context but never stated outright (eg. is it gene? genotype? allele?) many providers of such data have type-specific URLs and integration fails catastrophically (either collides undetected or 404s) if the wrong type is assumed. However, if wikidata is ingesting solely data that is already well documented with HTTP URIs, then I agree; start with the HTTP URI. 

However, let's first unpack what we mean by 'display' as it is rather loaded.

Scenario one: whatever users 'see' baked in the URLs of wikidata interface (or other 3rd party #researchparasites like us at Monarch). 
Scenario two: whatever is actually displayed most prominently in the 3rd party application web page itself
Scenario three: whatever is represented in the land of 3rd party API and graph stores (RDF)

For scenario one, the main objective is stability and identifiability, rather than immediate human recognition. You could bake in an http uri, for example: wikidata.org/record/uri=[http uri]. This works OK if: 
A) you're more-or-less in control of the data you're "consuming", and 
B) there's one primary URI
C) that URI is not volatile

eagle-i federated search uses this approach, for instance:

Inline image 1

However, 'control' and 'uncontested uri' and 'permanence' occur infrequently and they almost never co-occur. If 3rd party consumers/integrators don't have a well-planned strategy a-priori for to collapse distributed identifier equivalents, they may find it easy to aggregate, but difficult to integrate. Especially for highly-accessed resources (like chembl) with lots of relevant resources that are distributed and that reference the same chembl entity in endlessly different ways. 

For scenario two the main objective is human recognition. On this I'm for whatever works (labels, names, pictures, whatever), provided that anchor text is also linked to http URIs and at some point in the link chain transparent about its origins.

For scenario three the main consumer of that format is the tech elite. Here, the HTTP URI is a first class citizen, however even here, the self-documenting nature of formats like RDF/JSON is that CURIE identifiers can still be made easier on the human eye. Many Sporny has a candid and funny take on this. 

In PrefixCommons, we adopt a JSON-LD context approach (a cornerstone of schema.org that is already in use in over 10 million websites in the web-at-large). It it is efficient and scaleable for handling these distributed identifiers, especially for scenarios 1 and 3 above. 

This is the approach we use in Monarch. Wherever possible we display human-readable labels in the UI; however what we embed in our URLs are always the location-independent curie version of the ID. I've added to our identifier documentation marked up screenshots of our identifiers in action in our site using NCBIGene:6622 as an example. 
Inline image 2

I've omitted the painful parts about the state of the identifiers when we find them in ingested data, but just as a quick illustration, see this (38 distinct shortform representations, and 14 distinct http URIs). Note that all these permutations correspond to a single identifier, NCBIGene:6622, not to a class of identifiers, or a class of providers, or to similar entities spread across disparate resources. The state-of-affairs in database lookups is much the same as what you see in the web-at-large but to document that that would require a lot more effort.

The bottom line is that it doesn't matter how messy the identifiers are when you find them. Or how many different ways you find them. You can always create a curie form that links to the right place and you can use JSON-LD to provide context-specific direction. 

For instance: in the following contexts expand the CURIE to the chembl webpage, in these other contexts, expand to the (sometimes unresolvable) HTTP URI, in these other contexts, expand to the wikidata page for the entity. This approach is simple and it works.

Hope this was helpful?

Best wishes,

ps. The heterogeneity of identifiers also has major implications for data citation in journals; however, I've not touched that subject here as it is quite different (there's an archived PDF that can not be dynamically updated to change the identifiers or links).